[cxx-abi-dev] A mangling ambiguity

[Martin v. Löwis]

"J. Stephen Adamczyk" <jsa at edg.com> writes:

> template <class T> struct A {
>   template <class U> operator U() { return 0; }
> };
> int main() {
>   A<float> a;
>   int i = a;
> }
> 
> At the indicated point, the result type of the conversion function ("T_")
> should end, and the template argument list that follows ("IiE") should
> apply to the conversion function itself, not to its return type.
> There's no way for a demangler to know that

Why not? If this were a template argument to the conversion-type-id,
there would have to be another template argument list, which specifies
the conversion operator's template arguments, e.g. something like

template <class T> struct A {
    template <class U> operator U() { return sizeof(U); }
};


template <class T> struct B {
    template <template<class U>class X> operator X<double>() { return X<double>(); }
};

int main() {
  A<double> a;
  B<long> b;
  a = b;
}

Unfortunately, I can't try this out, since g++ rejects it. But I
believe the mangled name should be something like

_ZN1BIlEcvT_IdEI1AEEv


> This is illustrated by what the g++ demangler does with this name:
> 
> A<float>::operator float<int>()
> 
> This seems like a genuine ambiguity.  Or does someone see an out I'm
> missing?

The demangler is just confused: It doesn't recognise that it hasn't
even seen the template argument list, yet, when demangling the
T_. Instead, it should delay outputting the conversion-function-id
until it has the complete identifier (including its final template
argument list).

For that matter, it appears g++ is also confused. It compiles the
program

#include <stdio.h>

template <class T> struct A {
    template <class U> operator U() { return sizeof(U); }
};


int main() {
  A<double> a;
  int x = a.A<double>::operator int();
  double y = a.A<double>::operator double();
  int z = a.A<double>::operator double<int>();
  printf("%d %f %d\n", x, y, z);
}

Somebody please correct me if I'm wrong: A<double>::operator
double<int> is not a valid conversion-function-id in this case, is it?

Regards,
Martin